Challenges 2018 Megvii (Face++) Team lizeming@megvii.com I. COCO1 8 - PowerPoint PPT Presentation

MSCOCO Instance Segmentation Challenges 2018 Megvii (Face++) Team lizeming@megvii.com

I. COCO’1 8 Instance Seg Zeming LI Jian SUN Yueqing ZHUANG Xiangyu ZHANG Gang YU

Overview Improvements The results is obtained on test-dev Mask mmAP Detector mmAP 55 60 56.0 48.8 50 46.7 55 52.6 45 50 37.6 40 45 41.6 35 28.4 40 37.4 30 25 35 2015 2016 2017 Ours 2015 2016 2017(Megvii) Ours Object Detector Instance Segmentation 3.4% improvement 2.1% improvement

Outline 1) Location Sensitive Header 2) Backbone Improvement 3) Two-Pass Pipeline 4) Results

Outline 1) Location Sensitive Header 2) Backbone Improvement 3) Two-Pass Pipeline 4)Results

Mask RCNN Baseline FPN Original Mask Head Instance Seg Det mmAP mmAP Original Paper(detectron 1x) 33.6 - Our Re-implement 34.4 37.0

Location Sensitive Header Overall Architecture Comparison

Location Sensitive Header 1) Location Sensitive Detector name Mask AP Bbox AP Improvement Baseline 34.4 37.0 - + Local Sensitive Detector 35.4 38.7 + 1.0 / +1.7

Location Sensitive Header 2) Multi-Scale RoI name Mask AP Bbox AP Improvement Baseline 34.4 37.0 - + Local Sensitive Detector 35.6 38.7 + 1.0 / +1.7 + Multi-Scale RoI 35.8 38.9 + 0.2 / +0.2

Location Sensitive Header 3) Heavier Header name Mask AP Bbox AP Improvement Baseline 34.4 37.0 - Heavier Header 35.3 36.8 + 0.9 / -0.2

Location Sensitive Header 4) Mask Edge Loss name Mask AP Bbox AP Improvement Baseline 34.4 37.0 - Mask Edge Loss 35.0 37.0 + 0.6 / +0.0

Location Sensitive Header 4) Mask Edge Loss Sigmoid Cross Entropy

Location Sensitive Header Review of overall Architecture Location Sensitive Header: 1) Location Sensitive Detector 2) Multi-Scale RoI 3) Heavier Header 4) Mask Edge Loss

Location Sensitive Header Overall Performance in Small and Large Model BackBone Header Mask AP Bbox AP Improvement ResNet50 Baseline 34.4 37.0 - Location Sensitive Header 37.0 39.3 + 2.6 / + 2.0 ShuffleV2-GAP Baseline 40.3 45.0 - Location Sensitive Header 42.3 46.5 +2.0/+1.5 We will introduce backbone in next slides

Outline 1) Location Sensitive Header 2) Backbone Improvement 3) Two-Pass Pipeline 4)Results

Backbone Improvement 1. Channel Information Flow Ma N, Zhang X, Zheng H T, et al. ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design[J]. 2018.

Backbone Improvement 2. Add Global Information name Mask AP Bbox AP Improvement Baseline 34.4 37.0 - +GAP 35.1 37.7 +0.7/+ 0.7

Outline 1) Location Sensitive Header 2) Backbone Improvement 3) Two-Pass Pipeline 4)Results

Two-Pass Pipeline

Outline 1) Location Sensitive Header 2) Backbone Improvement 3) Two-Pass Pipeline 4)Results

Results Trained On Megvii’s Megbrain name Mask AP(val) Bbox AP(val) Improvement ResNet50 ( 2x-2batch-setting) 36.1 39.3 - ShuffleV2 (1batch) 40.3 45.0 +3.8/+5.7 2x Means 2x training setting used in Detectron

Results Trained On Megvii’s Megbrain name Mask AP(val) Bbox AP(val) Improvement ResNet50 ( 2x-2batch-setting) 36.1 39.3 - ShuffleV2 (1batch) 40.3 45.0 +3.8/+5.7 + Location Sensitive Header 42.3 46.5 +2.0 /+1.5

Results Trained On Megvii’s Megbrain name Mask AP(val) Bbox AP(val) Improvement ResNet50 ( 2x-2batch-setting) 36.1 39.3 - ShuffleV2 (1batch) 40.3 45.0 +3.8/+5.7 + Local Sensitive Header 42.3 46.5 +2.0 /+1.5 + 2 Batch Per GPU 44.5 49.3 +2.2/ 2.8 + Multi Scale Training + BN training

Results Trained On Megvii’s Megbrain name Mask AP(val) Bbox AP(val) Improvement ResNet50 ( 2x-2batch-setting) 36.1 39.3 - ShuffleV2 (1batch) 40.3 45.0 +3.8/+5.7 + Local Sensitive Header 42.3 46.5 +2.0 /+1.5 + 2 Batch Per GPU 44.5 49.3 +2.2/ 2.8 + Multi Scale Training + BN training + Improve on Dets 47.6 55.4 +3.1/ 6.1

Results Trained On Megvii’s Megbrain name Mask AP(val) Bbox AP(val) Improvement ResNet50 ( 2x-2batch-setting) 36.1 39.3 - ShuffleV2 (1batch) 40.3 45.0 +3.8/+5.7 + Local Sensitive Header 42.3 46.5 +2.0 /+1.5 + 2 Batch Per GPU 44.5 49.3 +2.2/ 2.8 + Multi Scale Training + BN training + Improve on Dets 47.6 55.4 +3.1/ 6.1 + Seg Multi-scale Testing 48.1 55.4 +0.5/0.0

Results Trained On Megvii’s Megbrain name Mask AP(val) Bbox AP(val) Improvement ResNet50 ( 2x-2batch-setting) 36.1 39.3 - ShuffleV2 (1batch) 40.3 45.0 +3.8/+5.7 + Local Sensitive Header 42.3 46.5 +2.0 /+1.5 + 2 Batch Per GPU 44.5 49.3 +2.2/ 2.8 + Multi Scale Training + BN training + Improve on Dets 47.6 55.4 +3.1/ 6.1 + Seg Multi-scale Testing 48.1/ 48.8(dev) 55.4/ 56.0(dev) +0.5/0.0 Instance Segmentation is obtained by single instance segmentation model

Results Trained On Megvii’s Megbrain name Bbox AP(val) Improvement Baseline 49.3 - +Soft-Nms 49.8 +0.5 +Multi-scale Testing 51.6 +1.8 +Ensemble 53.6 +2.0 add an additional model for ensemble: 55.4 +1.8 +with cascade R-CNN +external COCO++ 11W data

Results

Visualization Comparison Our baseline Location Sensitive Header Refine Location Error

Visualization Comparison Our Baseline Location Sensitive Header Refine Location Error

Visualization Comparison Our Baseline Location Sensitive Header Refine Location Error

Visualization Comparison Our Baseline Location Sensitive Header

Visualization Comparison Our Baseline Location Sensitive Header

Visualization Detector Results Mask Results

Visualization Mask Results Detector Results

Summary & thanks 1. Location Sensitive Header 2. Backbone Improvement 3. Pipeline Optimization Other Improvements: 1. Multi-Scale Training 2. Large Batch (MegDet : [C. Peng, CVPR’ 18]) 3. Multi-Scale and Flip Testing 4. Ensemble (only for Detection)

Looking for Interns, Researcher ， Research Engineer career@megvii.com yugang@megvii.com